Prompt Injection Gets Serious When Agents Can Run Commands on Your Endpoint
Aseel23_Folate · reddit · 2026-11-17
The author argues prompt injection has changed in nature: it's no longer just making a chatbot ignore instructions—agents now hold a local shell and access to your dev environment.
- Microsoft Defender's new AI Agent Runtime Protection preview inspects hooks inside the agent loop (prompts, tool calls, tool responses) and can audit or block risky activity before a command fires, instead of only checking final model output.
- Key scenario: if a coding agent reads a poisoned repo README or webpage and runs powershell as the user, normal MDE process telemetry just sees powershell running in a normal user context—the why behind the agent's decision is invisible.
- Another angle: Acalvio-style deception—planting fake skills, decoy MCP endpoints, and planted creds so an agent wandering off its normal workflow becomes a high-confidence tripwire.
- Overall shift: from "sanitize the prompt" to "control and monitor everything the prompt can reach."
The author asks testers whether they've flipped any rings to block mode yet or are still in audit.
More from coding & agent
- Celesto: CI for coding agents claims 90% cheaper builds and 25% faster runs — aniketmaurya · 2026-11-17
- Dad builds a pop-up poetry book for his kid with Qwen and Codex vibe coding — lxfater · 2026-11-17
- One mouse input setting turns LLM-written GTA5 aimbot from wild shaking into merely missing — zhuowei · 2026-11-17
- CI tip: keep everything in one .yml file, multi-file setups repeat builds and run slow — DanielLockyer · 2026-11-17
- Chrome team flags outdated Modern Web Guidance plugin in Claude Code marketplace, urges 1.0 reinstall — nicolascraske · 2026-11-17
- Sift open-sourced: Sentry founder's self-hosted LLM inbox manager with labels and rules — zeeg · 2026-11-17